The software factory · the overview

What is a software factory?

The phrase is everywhere in 2026, and it no longer means what it used to. Here's the actual machine people are describing — its four parts, where the human goes, and how close it really is.

"Software factory" is a recycled term. Two decades ago it meant Microsoft's model-driven toolkits, and later the defense-world delivery pipelines like the Air Force's Kessel Run. The 2026 version is different and more ambitious: an agentic software factory — a system you hand a specification, that then autonomously produces working, tested, deployed software, while humans set the intent and guard the quality bar instead of writing the code. The name is the claim: factories are machines that build things, and this is a machine that builds software.

Strip away the marketing and almost every serious description converges on the same four-part machine. Click through it:

Where the human goes

The interesting move isn't that the machine writes code — it's what it does to the job. In a working factory the human stops being the person who types functions and becomes two things instead: the one who writes a specification precise enough to act on, and the one who decides whether to trust the result. That's not a smaller job; it's a different and arguably harder one. You don't get there in one leap either — teams climb a trust ladder, from agents that can only read, to agents that draft pull requests, to agents that merge on their own inside guardrails. We break that ladder into concrete stages in the five levels of agent autonomy.

Where it actually is

The headlines are breathless — fleets of unattended agents at big-name companies merging hundreds of pull requests a week, engineers who claim they haven't hand-written code in months. Some of that is real and some is a pitch deck. The honest picture from 2026's own numbers is messier: most teams adopting agent fleets hit a wall not on generation but on trust. Surveys show reliability and hallucination management as the top adoption barrier, developer confidence in AI-written code actually falling year over year, and — the punchline — review becoming the new bottleneck: far more pull requests, far larger diffs, and review time growing faster than anyone shipped it. The factory can generate all day; the constraint is the human gate at the end. That gate is the whole reason code review is the software factory's real bottleneck — and it's the least glamorous, most decisive part of the machine.

Why the metaphor is worth taking seriously anyway

Skepticism about the hype is healthy; dismissing the shape is not. A factory is a useful frame because it forces the right questions: what's the intake, who owns the spec, how do parallel workers avoid stepping on each other, what's the quality gate, and how does the line learn from defects. Even if full autonomy stays years out, teams that think in these four subsystems ship better with agents than teams that treat an agent as a faster autocomplete. The rest of this series takes the machine apart one subsystem at a time — starting with the two that decide whether any of it works: the specification going in, and the review gate on the way out.